rLLM / DeepScaleR: Agentic RL framework (Agentica)

Sijun Tan, Michael Luo, Colin Cai, Tarun Venkat, Kyle Montgomery, Aaron Hao, Tianhao Wu, Arnav Balyan, Manan Roongta, Chenguang Wang, Li Erran Li, Raluca Ada Popa, Ion Stoica

Open on GitHub

Type: repo
Venue: Agentica / Berkeley Sky Computing Lab (GitHub rllm-org)
Year: 2026
Source: github
Access: free
Language: en
Added: 2026-08-14T20:50:00Z
Verified: 2026-08-14T20:50:00Z

Summary

agentica-project/deepscaler redirected to rllm-org/rllm. Original DeepScaleR work (Feb 2025 blog) scaled RL on DeepSeek-R1-Distill-Qwen-1.5B using ~40k unique math Q-A pairs (AIME 1984-2023, AMC pre-2023, Omni-MATH, Still). Current rLLM is a general agentic RL stack: GRPO/REINFORCE/RLOO/SFT/on-policy distillation; backends verl, tinker, fireworks; 60+ evals. Later Agentica results include DeepCoder-14B and DeepSWE-32B. Discord posted the old deepscaler/verl tree.

Keywords

deepscaler · rllm · agentica · grpo · math-rl · verl · berkeley-sky

Topics

reinforcement learning, LLM post-training, math reasoning, agentic RL

Research notes

Primary: GitHub README + API (Apache-2.0). Discord posted https://github.com/agentica-project/deepscaler/tree/main/verl which now 301s to rllm-org/rllm. Blog https://pretty-radio-b75.notion.site/DeepScaleR-Surpassing-O1-Preview-with-a-1-5B-Model-by-Scaling-RL-19681902c1468005bed8ca303013a4e2. Preview dataset/model not in datasets_local as themselves (only uiuc-kang-lab noisy-label derivatives tied to papers_local 481). Not previously cataloged as DeepScaleR paper/repo.